Skip to content

An agentic browser harness: an AI side panel & orb in your own logged-in Chrome, Edge, Arc, Brave or even Firefox run by the agent/subscription you already use (Claude Code, Codex, Antigravity). Free, open source, no API key. Also an MCP server.

feat(mcp): list every routable skill from the CLI

Shahriar7afa371c8d24
diff --git a/mcp/src/cli.ts b/mcp/src/cli.ts
index 36dcc60..d6ea9f3 100644
--- a/mcp/src/cli.ts
+++ b/mcp/src/cli.ts
@@ -2,6 +2,8 @@ import { readFileSync } from 'node:fs';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
import { describeActions } from '@/lib/actions/registry';
import { assertToolNamesRoundTrip, toolNameFor } from '@/lib/actions/tool-names';
+import { join } from 'node:path';
+import { loadSkills, skillDirNames, uploadedSkillsDir } from './agent/skills';
import { ensureDaemon, probeExisting } from './ensure-daemon';
import { clearLockfile, logPath, readLockfile } from './lockfile';
import { log } from './log';
@@ -16,6 +18,7 @@ const USAGE = `voicelink-mcp ${pkg.version} — MCP access to your browser via t
voicelink-mcp sessions list paired browsers
voicelink-mcp revoke [origin] unpair one browser, or all of them
voicelink-mcp tools print the bundled tool manifest (no browser needed)
+ voicelink-mcp skills list the skills the agent can route to, and where they came from
voicelink-mcp status show daemon and extension connection state
voicelink-mcp stop stop the background daemon
voicelink-mcp logs print the daemon log
@@ -53,6 +56,9 @@ switch (command) {
case 'stop':
stop();
break;
+ case 'skills':
+ printSkills();
+ break;
case 'logs':
showLogs();
break;
@@ -79,7 +85,8 @@ async function serve(): Promise<void> {
// Present when the daemon spawned a Claude Code run and that run spawned us: tagging the
// invocations is what routes them through the run's approval gate and timeline.
const bridge = await RemoteBridge.connect(lock.port, lock.token, process.env.VOICELINK_AGENT_RUN);
- const server = createMcpServer(bridge, pkg.version);
+ // Only a spawned agent run gets the map-writing tool; an ordinary MCP client never sees it.
+ const server = createMcpServer(bridge, pkg.version, { agentRun: !!process.env.VOICELINK_AGENT_RUN });
await server.connect(new StdioServerTransport());
log(`stdio MCP server attached to daemon on port ${lock.port}`);
@@ -108,6 +115,34 @@ function printTools(): void {
);
}
+/**
+ * Every skill the router can reach, in precedence order — the answer to "did my upload land,
+ * and is it shadowing something?". Reads the same directories a run does, so it needs no daemon.
+ *
+ * It also puts the skill loader in this bundle's import graph, which is what makes
+ * `npm run mcp:manifest` catch a browser-only import creeping into `lib/skills/`.
+ */
+function printSkills(): void {
+ const skills = loadSkills();
+ if (!skills.length) {
+ console.log(`No skills found. Looked in:\n ${skillDirNames().join('\n ')}`);
+ return;
+ }
+ for (const skill of skills) {
+ const scope = skill.category === 'site-exploration' ? skill.domains.join(', ') || 'no domains — @name only' : '';
+ const tags = [
+ skill.source,
+ skill.provenance === 'generated' ? 'mapped' : skill.category,
+ scope,
+ skill.isDefault ? 'default' : '',
+ ].filter(Boolean);
+ console.log(`${skill.name} (${tags.join(' · ')})`);
+ if (skill.description) console.log(` ${skill.description}`);
+ if (skill.provenance === 'generated') console.log(` ${join(uploadedSkillsDir(), skill.name)}/`);
+ }
+ console.log(`\nRead in order: ${skillDirNames().join(' → ')} (a later one shadows an earlier one by name)`);
+}
+
async function showStatus(): Promise<void> {
const lock = await probeExisting();
if (!lock) {